FINANCIAL NEWS SENTIMENT ANALYSIS USING MODIFIED VADER FOR STOCK PRICE PREDICTION

• Main Author: Prameswari Ekaputri, Anindya
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/66312
• Institution: Institut Teknologi Bandung

There are many factors to influence stock price movements, and one of those factors is opinions about the company. This paper builds a model to analyze sentiments in news about a company, then uses the sentiment scores as a feature together with social media sentiments and historical stock price to predict stock prices in the future. Analysis and prediction were carried out on four Indonesian banks: BBNI, BBRI, BMRI, and BBCA. Sentiment analysis model were built by modifying VADER, a rule-based model that uses lexicons, then experiments were conducted by comparing performances between general lexicon, financial lexicon, and combined general-financial lexicon. The best result from the experiment is averaged per day to transform the data into time-series. It is then used together with social media sentiments and historical stock price to build a price prediction model. Experiments for prediction model were conducted in two steps: the first step aims to determine the range of days that needs to be considered toget optimal prediction results; the second step aims to determine the best model architecture and feature combinations. In the first step, several ranges of days were experimented; the range that produced the best results were reused in the second step. In the second step, performances of random forest, LSTM, and CNN models were compared by varying the number of layers and units/filters along with varying features (with or without sentiment). Experiment results shows that sentiment analysis model with financial lexicon performs best with 73% accuracy, while prediction models perform best when trained without sentiment features. In general, CNN models give better results with MAPE values ranging from 1.5% until 2.8%